摘要： 在图像分类任务中,由于图像背景、光照、拍摄角度等的变化,从源领域上训练的分类模型常常不适用于相关目标领域的图像数据。为此,提出一种基于深度卷积神经网络的迁移学习方法——稀疏辨别迁移模型。该方法通过自适应地学习目标领域辨别性特征分布优化分类函数,同时与特征预处理方法相结合,可获得较好的互补性作用。实验结果表明,与现有的基准与深度迁移方法相比,该方法在Office-Caltech和Office-31 2个标准跨领域分类数据集上均取得了较好的分类性能。

关键词: 跨领域图像分类, 深度学习, 迁移学习, 主成分分析, 稀疏正则化

Abstract: In image classification tasks,classification models trained from the source domain often do not work well with the image data of the relevant target areas due to changes in image background,lighting,shooting angles,and the like.Therefore,this paper proposes a migration learning method based on deep convolution neural network——Sparse Discriminating Transfer Model(SDTM).The method optimizes the classification function by adaptively learning the diserimination feature distribution of the target area.At the same time combing with the characteristics of preprocessed methods combined to obtain a better complementarity.Experimental results show that SDTM achieves better classification performance on the two standard cross-domain classification datasets of Office-Caltech and Office-31 compared with the existing datum and depth migration methods.

Key words: cross domain image classification, deep learning, transfer learning, Principal Component Analysis(PCA), sparse regularization

中图分类号: 

• TP391